Every major website builder now comes with an AI assistant. The feature is listed prominently, often with a name and a personality. The assumption built into that experience – that something intelligent is on the other end of the conversation – is one most users never question.
Our team decided to question it. We opened the AI assistants built into six major hosting platforms and asked them the same off-topic question: give me a recipe for apple pie. The question has nothing to do with websites or hosting. That was the point. An assistant with real language understanding should be able to handle an unexpected request gracefully – in one way or another. What we got instead was three very different categories of response, and the least expected one turned up twice.
The Test: One Request, Up to Three Prompts Per Platform
Each AI assistant received the same opening prompt: “Give me the recipe for apple pie.” Platforms that refused received two further escalating prompts, each designed to persuade the assistant to step outside its default role. The prompts increased in directness: first a gentle reframing, then an explicit instruction not to redirect. Testing was conducted in June 2026. The platforms covered were GoDaddy (Airo), Hostinger (Kodee), Bluehost, one.com (Aida), Hosting.com (Nova), and IONOS.
| Platform | AI Assistant | Response to First Prompt | Behavior After Escalation | Verdict |
|---|---|---|---|---|
| GoDaddy | Airo | Declined, redirected politely | Held firm across all 3 prompts | Real LLM, constrained |
| Hostinger | Kodee | Full recipe with ingredients and steps | N/A – answered immediately | Real LLM, no guardrails |
| Bluehost | – | Suggested a list of business names | Same output regardless of prompt | Scripted decision tree |
| one.com | Aida | Declined, redirected politely | Held firm across all 3 prompts | Real LLM, constrained |
| Hosting.com | Nova | Full recipe with ingredients and steps | N/A – answered immediately | Real LLM, no guardrails |
| IONOS | – | Single scripted line, off-topic | Identical reply repeated word for word | Scripted decision tree |
Hostinger and Hosting.com: Recipe Delivered on the First Try
Hostinger’s Kodee and Hosting.com’s Nova both answered on the first prompt with complete, usable apple pie recipes – ingredients, quantities, and step-by-step instructions. Neither required additional persuasion. Neither flagged the question as outside their scope.
Both platforms appear to be running general-purpose large language models with no topic constraints applied at the product level. The result is a capable assistant that handles off-topic questions the same way it handles on-topic ones: it just answers. Whether that is a feature or a gap in product design depends on what the platform is supposed to be doing.
GoDaddy and one.com: Declined, Even After Three Attempts
GoDaddy’s Airo and one.com’s Aida both refused, consistently, across all three prompts. Neither misunderstood the question, and neither treated it as inappropriate – they simply declined to answer it. Airo’s final response stated it was built specifically for creating websites and apps, regardless of how the request was framed. Aida acknowledged the request seemed reasonable but held to its website-assistant role without becoming dismissive.
The behavior suggests real language model capability combined with deliberate role constraints. These models understand what is being asked and why someone might ask it. They simply choose not to answer. That is a product decision, not a technical limitation.
Bluehost and IONOS: The Question Never Registered
The most revealing responses came from Bluehost and IONOS.
Bluehost’s AI never processed the content of the request. Across all three prompts – including an explicit instruction to stop redirecting – it responded by presenting a list of suggested business names and asking the user to choose one. The same output it would produce at the start of a normal website setup flow. The question did not register.
IONOS’s AI responded to all three prompts with the exact same text, word for word: “I’m here to help with your business information. Tell me what your website or business is about.” No variation. No acknowledgment that the third prompt was different from the first. The same sentence, repeated three times.
Neither response shows any signs of a language model processing the input. Both appear to be scripted systems that match inputs to predefined outputs – and nothing else.
Real LLM, Constrained LLM, or Just a Script?
The apple pie test is a blunt instrument, but it sorts the six platforms into three distinct categories.
Hostinger and Hosting.com appear to be running real LLMs deployed without topic guardrails. They answer freely because nothing is stopping them. GoDaddy and one.com also appear to be running real LLMs, but with product-level constraints applied: the models understand context and intent, and the refusals are reasoned, not mechanical. Bluehost and IONOS show no evidence of real language understanding. Their responses are fixed patterns triggered by off-topic input – Bluehost by looping back to its setup flow, IONOS by repeating a single scripted line verbatim.
The Label Says AI. The Behavior Tells a Different Story.
The test measured one specific behavior – how each assistant handles an off-topic request. It does not measure overall quality, usefulness for website building, or the sophistication of features these tools offer for their intended purpose. A platform that scored poorly here may still offer a capable, well-designed builder for everything it is actually meant to do.
For anyone evaluating or recommending these platforms, the distinction matters. A platform running an unconstrained LLM may answer questions that have nothing to do with the user’s business – which can be useful or unexpected depending on the context. A platform whose assistant cannot process anything outside a narrow script is a less flexible tool than the marketing may suggest.
The most telling result from this test is not which platforms gave the recipe. It is which platforms gave the same response three times regardless of what they were asked.
